Onboarding a new hire when the AI already knows the company

There's a version of new hire onboarding every small company knows by feel. Three weeks, roughly. Week one, the new hire reads whatever documentation exists and asks where things are. Week two, they produce their first drafts, which come back covered in corrections. Week three, the corrections taper, and somewhere in there the tribal knowledge finishes soaking in, the knowledge nobody wrote down because nobody knew they were carrying it.

I've now watched first weeks on teams where the ramp runs differently, because the sayable half of that tribal knowledge already lives in the team's AI tools. The setup is simple to describe. These teams run Company Brain, a shared memory of approved company facts that loads into every AI session through MCP. Everyone, including a new hire on day one, connects Claude, ChatGPT, Copilot or Cursor to the same shared address, and changes to the facts arrive as proposals that a curator reviews. The AI a new hire opens on her first afternoon already knows the company.

What follows is the diary of those two weeks. I've seen enough of them now to write it in advance. Yours may differ in the details. I doubt it differs much.

Day one: the drafts sound like the team by the afternoon

Her first real task is a draft, because first tasks usually are. A customer email, a help page, an internal summary. She opens the AI she already knows from her last job, connected to the company's address that morning, and it spells the product names correctly, capitalization included. It knows the tone rules, the terms the team uses and the terms it deliberately avoids, and which claims about the product are approved. Her draft comes back from review with comments about substance.

Compare that with the traditional first draft, which is competent and wrong in twenty small ways. A product name lowercased. A package described by its old name from two rebrands ago. A tone pitched for a company three times more formal. None of those errors is serious, and every one of them normally becomes a correction from a colleague, because corrections were how the knowledge traveled. It worked, at the speed of embarrassment.

The detail that gets me every time is how little anyone prepared. No AI onboarding guide, no curated prompt folder for the new person. She connected her tools and inherited the same context as everyone else, because there is only one context to inherit.

The questions change by day four

Week one questions are traditionally vocabulary questions. What do we call the smaller package. Do we say clients or customers. Is that integration officially supported or just tolerated. These are the cheapest questions to answer and the most expensive to keep answering, because each one lands as an interruption on whoever sits closest.

On these teams, the vocabulary questions go to the AI, and it answers correctly because the answers were written down as facts and reviewed by someone who knew. What the new hire asks her colleagues instead are judgment questions. Why did we set the boundary there. This customer is pushing on the policy, what do we actually weigh. The tone rules say plain, and this apology reads too plain, where's the line. These are questions worth a colleague's time. They're also the questions that teach the company, and she gets to them in week one instead of week three.

The walking FAQ gets her afternoons back

Every small company has one. The person who's been there longest and carries the most unwritten knowledge, so every new hire orbits them for a month. Onboarding cost is usually counted as the new hire's ramp time. The veteran's lost weeks rarely make the ledger, and they're often the more expensive line, because the veteran is the person the company can least afford to slow down.

When the sayable knowledge sits in the shared memory, the orbit shrinks. The veteran still onboards the new hire, on the parts only a human can transfer, and we'll get to those. What disappears from her calendar is the FAQ layer: the spelling corrections, the "we renamed that in the spring", the fifth retelling of what the standard package includes. Watching this happen taught me who the shared memory is really for. The new hire benefits on day one. The veteran benefits every day after.

Where the three weeks were actually going

The traditional ramp is mostly waiting for context to arrive through slow channels. Shadowing, corrections, overheard calls, an old thread found by luck. The knowledge existed on day one. It lived in heads and scattered documents, so delivery took weeks and depended on collisions.

Teams have always tried to fix this with a shared context document, and it genuinely helps until the document forks and quietly rots. I've written the biography of that document and why it stops scaling at three people. The structural fix keeps the discipline of writing things down and repairs the delivery. Facts that are short, owned and reviewed, loaded into every AI session automatically, reach the newest hire in exactly the state the curator left them. There is no pasting ritual for her to learn and no stale copy for her to inherit.

What the facts cannot carry, and why that's fine

Now the honest part. Facts hold the sayable: names, terms, rules, policies, positioning, anything that fits in one reviewable sentence. Judgment does not compress that way. Neither does taste, or the history behind a decision, or knowing which customer needs a phone call and which needs a week of quiet. That knowledge still transfers the way it always has, human to human, over weeks, through real work reviewed by someone who cares how it turns out.

I'd go further: that transfer is the point of onboarding. Nobody founded a company hoping to spend hire number nine's first month on spelling. When the sayable layer arrives on day one through the tools, the human layer finally gets the attention it deserved all along. The veteran's freed afternoons turn into the conversations that make someone actually part of the team, which is the outcome the three-week ramp was always fumbling toward.

If your next hire starts in a month

The preparation is smaller than it sounds. Write down the facts you already correct in people's drafts every week. I've described a practical method for writing your first 50 company facts if you want somewhere to start. Give one person ownership of the review, because facts nobody owns rot like documents nobody owns. Connect the team's AI tools once, through the one shared address. Company Brain's free plan stays free, so running this experiment before a start date costs a curator's afternoon and nothing else.

Then pay attention to what your new hire asks in week one. If the questions are about judgment, the ramp is working.